ML Model Data Masking and Confidence Withholding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models often 'memorize' sensitive information from training data, leading to potential exposure and reverse engineering, which is undesirable for privacy and security reasons.
Innovation Solution
Abstracting and masking sensitive data before training machine learning models, and using confidence levels to determine when to withhold results to prevent exposure of sensitive information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sensitive information is included in training data to improve model accuracy, then model performance is improved, but privacy security deteriorates due to potential data exposure and reverse engineering
Solution Approach 1:
The patent extracts and removes sensitive information from the training data before model training. This is achieved through data cleaning processes that identify and filter out personally identifiable information (PII), financial data, and other sensitive content, thereby preventing the model from memorizing and exposing such information while maintaining the useful patterns needed for accurate predictions
Solution Approach 2:
The patent introduces an intermediary data processing layer between data collection and model training. This intermediary layer includes data anonymization, aggregation, and transformation processes that convert raw sensitive data into sanitized training data, allowing the model to learn from data patterns without direct exposure to sensitive information
2Object-affected harmful factors
If data is cleaned and sanitized before training to prevent sensitive information exposure, then privacy security is improved, but data processing complexity increases
Solution Approach 1:
The patent performs data cleaning, sanitization, and anonymization actions before the model training process begins. By conducting these privacy-preserving operations in advance during data preparation, the system avoids the need for complex post-training filtering and explanation mechanisms, thereby managing processing complexity through upfront data transformation
Solution Approach 2:
The patent divides the data processing pipeline into distinct sequential stages: data collection, data cleaning, data sanitization, data anonymization, and model training. This segmentation allows each stage to focus on specific tasks, making the overall complex process more manageable and efficient through modular processing
3Adaptability or versatility
If the model is trained on diverse and comprehensive data to improve learning capability, then model versatility is improved, but the risk of memorizing and exposing sensitive information increases
Solution Approach 1:
The patent applies different processing quality levels to different parts of the training data. High-quality detailed processing is applied to data segments that contain sensitive information, while standard processing is applied to other segments. This allows the model to learn from diverse data comprehensively while providing enhanced privacy protection where needed
Data Source
AI summary
Data may be abstracted and/or masked prior to being provided to a machine learning model for training. A machine learning model may provide a confidence level associated with a result. If the confidence level is too high, the machine learning model or an application including the machine learning model may refrain from providing the result as an output. In some examples, the machine learning model may provide a “second best” result that has an acceptable confidence level. In other examples, an error signal may be provided as the output. In accordance with examples of the present disclosure, data may be abstracted and/or masked prior to being provided to a machine learning model for training and confidence levels of results of the trained machine learning model may be used to determine when a result should be withheld.